Presentation is loading. Please wait.

Presentation is loading. Please wait.

Centre de Calcul de l’Institut National de Physique Nucléaire et de Physique des Particules Data storage services at CC-IN2P3 Jean-Yves Nief.

Similar presentations


Presentation on theme: "Centre de Calcul de l’Institut National de Physique Nucléaire et de Physique des Particules Data storage services at CC-IN2P3 Jean-Yves Nief."— Presentation transcript:

1 Centre de Calcul de l’Institut National de Physique Nucléaire et de Physique des Particules Data storage services at CC-IN2P3 Jean-Yves Nief

2 Storage at CC-IN2P3: disk FJPPL Computing Workshop05-03-2015 2 Parallel File System: GPFS (1.9 PBs) File servers: xrootd, dCache (11 PBs) Used for High Energy Physics (LHC etc…) Mass Storage System: HPSS (600 TBs) Used as a disk cache in front of the tapes. Middlewares: SRM, iRODS (840 TBs) Databases: mySQL, PostGres, Oracle (57 TBs) Software Direct Attached Storage servers (DAS): Dell servers (R720xd + MD1200) ~ 240 servers Capacity: 12 PBs Disk attached via SAS: Dell servers ( R620 + MD3260) Capacity: 1.9 PBs Storage Area Network disk arrays (SAN): IBM V7000 and DCS3700, Hitachi HUS 130. Capacity: 240 TBs Hardware

3 Storage at CC-IN2P3: tapes FJPPL Computing Workshop05-03-2015 3 Mass Storage System: HPSS 24 PBs Max traffic (from HPSS): 100 TBs / day Interfaced with our disk services Backup service: TSM (1 PB) 4 Oracle/STK SL8500 librairies: 40,000 slots (T10K and LTO4) Max capacity: 320 PBs (with T10KD tapes) 106 tape drives 1 IBM TS3500 library: 3500 slots (LTO6) Hardware Software

4  4 Oracle/STK SL8500 libraries (4 x 10000 slots) for HPSS and TSM.  1 IBM TS3500 library (2731 slots) for TSM.  Tape media: Storage hardware as of today FJPPL Computing Workshop05-03-2015 4 MediaCapacity Max Bandwidth Number of drives Service T10K-B1TB120 MB/s64HPSS T10K-C5 TB240 MB/s22HPSS T10K-D8.5 TB252 MB/s20HPSS LTO4800 GB120 MB/s20TSM LTO62.4 TB160 MB/s6TSM

5 How experiments are dealing with the storage ? FJPPL Computing Workshop05-03-2015 5 TSM HPSS iRODS SRM Client apps (local or distant apps) GPFS dCache xrootd File servers / filesystem (local access) Middlewares (wide area access) Mass Storage System Backup

6  Purpose: ◦ For heavy I/O operations by many batch jobs in //. ◦ Not a permanent repository.  60 servers, 1.9 PBs of disk space.  1000 GPFS clients installed on our  Used by: ◦ up to 25% of the batch farm. ◦ 84 groups.  Max activity observed: 500 TBs read in 5 days (Planck)  To do: ◦ Migration of metadata storage to flash technology (NetApp EF560). ◦ Reduction of the number of space (35 right now) down to 4. Local access: distributed filesystem FJPPL Computing Workshop05-03-2015 6

7  Purpose: ◦ Data access by batch jobs (local + grid). ◦ Pattern: remote access (random I/O, sequential). ◦ Interface with HPSS:  Using Treqs (Tape request scheduler). ◦ dCache service part of the Atlas and CMS federations:  xrootd proxy servers (gateway to our local dCache service). ◦ dCache head nodes db: use of SSD.  To do: ◦ xrootd: lots of tape staging (eg: 20k requests per day)  « automatic » prestaging for some large « productions » needed. Local access: dCache-xrootd FJPPL Computing Workshop05-03-2015 7 service# serverscapacityexperiments # clients in // access protocol dCache1658.5 PiB LCG + « EGEE » VOs (17) 15000 dcap, xrootd, WebDAV, gsiftp xrootd442.5 PiB Alice + HEP + astroparticle (13) 8000 xrootd, http

8 Local access (xrootd/dCache) FJPPL Computing Workshop05-03-2015 8 Client HPSS Redirector server: Xrootd T1.root (1) (4) (2) /hpss/in2p3.fr/T1.root ? (etc…) Data server: Xrootd (5) (3) (1) + (2): load balancing (6): random access Data server: Xrootd (6) redirectors Data servers (4) + (5): dynamic staging (TreqS)

9  Purpose: ◦ Interface to our back end storage: tapes (main storage).  29 servers (disk + tape movers).  Huge activity: ◦ 23.3 PiB on tapes, 597 TiB of disk space (Dell R510, R720 servers). ◦ Up to 100 TiB in 22h between HPSS and other storage services.  New drive technology every 30 months.  Life cycle of a tape generation 8 years: ◦ Ramping up in 4 years. ◦ 4 more years in production.  T10KB decommissionning: ◦ 20000 tapes to be repacked until end 2017  5 PiB / year. ◦ Represents:  1000 T10K-T2  T10K-C  Or 750 T10K-T2  T10K-D  Tape lifetime: avg 7 years (max 10 years).  Redefine of the class of services.  Refactoring of Treqs (Tape request scheduler) by our dev team. Mass Storage service: HPSS FJPPL Computing Workshop05-03-2015 9

10  Dynamic staging may result in: ◦ Massive staging requests. ◦ Long delays to access the files. ◦ Tapes capacity is getting bigger!  Increased job inefficiency.  To leverage these drawbacks: ◦ Big files on tapes (several GBs). ◦ Reordering of file requests per tape id (reduce # of mounts/dismounts) ◦ Prestaging. ◦ Avoid file scattering on a large amount of tapes (tape families ?).  Refactoring of Treqs (Tape request scheduler) by our dev team. Local access: pitfalls and lessons learned FJPPL Computing Workshop05-03-2015 10

11  Purpose: ◦ Data management (storage virtualization, policies). ◦ Data sharing, access, archival etc… in a distributed and heterogeneous environment.  SRM: ◦ Used by Atlas, CMS, LHCb.  iRODS:  15 servers (DAS), 804 TBs of dedicated disk space.  Managing > 9 PBs of data (HEP, astro, biology, Arts & Humanities): includes data stored on tapes.  Software contributor.  Web interfaces, HTTP/REST interfaces: truly needed.  To do: simple data transfer tool replacement to bbftp. Middleware: SRM, iRODS FJPPL Computing Workshop05-03-2015 11

12 Future storage needs FJPPL Computing Workshop05-03-2015 12 Cumulated Disk (TB) Cumulated Tape (TB) Credit: Rachid Lemrani Hypothesis: LCG: +10% / year HEP like experiment: ratio disk/tape = 30% CTA not included

13  Tape will still have a major role.  Hard drives still preferred to SSD: ◦ prices will converge in the next couple of years. ◦ SSD usage limited at the moment for key apps (GPFS metadata, HPSS and dCache databases etc…).  Assessment of object storage: ◦ Swift: testing it at the moment. ◦ Swift extra value wrt what we already have: to be proven. ◦ Ceph (but it can be more than an object storage).  Qserv service for LSST (distributed database).  Hadoop: no plans yet as not much needs expressed by our user community.  Metadata: will increase massively with increased storage. ◦ Limit the amount of files => increased file size.  Data life cycle improvements: archival service ? Storage prospects FJPPL Computing Workshop05-03-2015 13


Download ppt "Centre de Calcul de l’Institut National de Physique Nucléaire et de Physique des Particules Data storage services at CC-IN2P3 Jean-Yves Nief."

Similar presentations


Ads by Google